What is Intelligence in the Context of Artificial

							               What is Intelligence in the Context of Artificial Intelligence?

                                           Adam Beautement.
                                         themorat@hotmail.com

                                  Themes and Debates Essay May 2000.
                               Nottingham University Psychology Department


INTRODUCTION

1.      The attributes required by an entity such that it can be labelled „intelligent‟ are the
subject of much debate in psychology and philosophy alike. The driving force behind artificial
intelligence (A.I.) is the desire to create an intelligent agent outside the realms of nature and
natural evolution. However as intelligence is poorly understood in the wider sense it seems
unproductive to create a distinction between that intelligence embodied by humans and
other biological agents, and any semblance of intelligence invested in artificial devices and
machines. Intelligence is universal, the criteria used to award the label „intelligent‟ (whatever
they may be) can be applied to any system regardless of form and construction. Artificial
intelligence is merely intelligence made manifest in a manufactured form.

2.      As there is no real sub-category of artificial intelligence as opposed to full intelligence
when we ask “What is intelligence in the context of Artificial Intelligence?” we are really
asking two questions, “What is intelligence?” and “What makes an artificial system
intelligent?” These are the questions that this essay will attempt to answer. During this
discussion ideas that are beyond the grasp of contemporary technology may be considered
as we are looking at the theoretical possibility of such devices not their physical
implementation. The notion of machine intelligence and its technological creation are
separate issues. We can conceptually conceive of a thinking artefact while lacking the
technology to construct it.

3.      The second of the questions asked above “what makes an artificial system
intelligent?” can be answered simply. An artificial system must suffer the same criteria as
any other system, so whatever makes a natural entity intelligent is the same thing that will
make an artificial system intelligent. However, this reply merely begs more questions and so
we are led to the first and more salient of the questions above, “What is intelligence?” In this
essay I will make some attempt to establish what comprises intelligence and I will then
consider whether or not intelligence can be artificially created.

4.      There is no real way to formally define intelligence. The first place to look for answers
is the natural world and the animal kingdom. It is here that we may find the only known
examples of intelligent and rational behaviour. Copeland discusses the attribution of
intelligence to the natural world in the following way:

        "The concept of a thinking thing is an inexact one, in that we have no precise
        criterion of how adaptable the behaviour-producing inner processes of an
        entity must be in order for it to qualify as a thinking thing. Going down the
        evolutionary scale there is an extremely fuzzy boundary between organisms
        that clearly do think and those that clearly do not. Chimpanzees do and
        maggots do not, but no one is sure what to say about crabs and salamanders.
        (Copeland 1993)."


(c) Adam Beautement May 2000                        1
5.       So how can we make a distinction? Where is it sensible to create a divide with
intelligent beings on one side and those without intelligence on the other? Certain extremist
views such as solipsism hold that only oneself and one‟s own experiences exist. Solipsism is
the extreme consequence of believing that knowledge must be founded on inner, personal
states of experience, (Blackburn 1994). If this view is taken then only by being the machine
in question can we ascertain whether or not it thinks. However it is then impossible to inform
anyone else of this information. Therefore, for the purpose of this discussion, it is assumed
that it is possible to ascertain whether a being separate to oneself can hold intelligence.


THE DISCUSSION

6.      Establishing where intelligence starts and stops cannot be answered by a survey or
statistical measure, it is distinctly non-empirical. However if intelligence exists as a definable
context then it can be tested. One of the more famous tests of intelligence in this field is the
Turing test. Although designed to assess machine intelligence it provides a useful basis for a
wider discussion. By looking at its strong points and its failings many of the issues relevant to
intelligence can be brought to light.

7.      Proposed by Alan M. Turing in his 1950 paper “Computing machinery and
Intelligence,” the test was originally described in terms of a game called the imitation game.
To conduct the test we need two people and the machine to be evaluated. One person plays
the role of the interrogator, who is in a separate room from the computer and the other
person. The interrogator can ask questions of either the person or the computer by typing
questions and receiving typed responses. However the interrogator knows them only as A
and B and aims to determine which is the person and which is the machine. The goal of the
machine is to fool the interrogator into believing that it is the person. If the machine succeeds
at this then we will conclude that the machine can think. (Rich and Knight 1991). In short it
should be impossible to distinguish between its conversation and the replies of a human.
There have been many attempts at creating devices and programs that can pass the Turing
test, and some have even been mistaken as human. However in the end they have all been
exposed because they failed to maintain the illusion of intelligence. And in all of these cases
it has been an illusion of intelligence for, despite being equipped with some useful tricks,
none of the programs have embodied intelligence.

8.     Two of the more successful programs to attempt the test were called Eliza, and Parry.
It has been claimed that Eliza has already passed the Turing test. An office worker trying to
contact one of his programming staff was connected to Eliza instead and the following
conversation ensued:


        Executive: Do you think that I could use the computer this morning?
        Eliza: WHY DO YOU ASK?
        Executive: I might be able to make some additional Telcomp sales.
        Eliza: WHY AREN‟T YOU SURE?
        Executive: My prospective customers have never seen a demonstration of the
        system.
        Eliza: DOES THIS MEAN VERY MUCH TO YOU?
        Executive: Of course it does.

(c) Adam Beautement May 2000                    2
        Eliza: YOU SEEM VERY POSITIVE ABOUT THAT.
        Executive: This is ridiculous.
        Eliza: WHAT DO YOU MEAN RIDICULOUS?
        Executive: Is there any reason why I can‟t use the computer now?
        Eliza: TELL ME MORE ABOUT YOUR PROSPECTIVE CUSTOMERS.

9.    After this exchange the real programmer received an irate phone call from the
executive demanding an explanation for his outrageous behaviour, (Copeland 1993).
However this exchange did not take place under the conditions specified by the test. There
was only one interrogator and the executive was not trying to distinguish between a human
and a machine. He began the conversation assuming he was dealing with another human
and therefore would have anthropomorphised the answers he was receiving.

10.     The tendency to anthropomorphise the results of computer programs further clouds
the issue of trying to establish what is and is not intelligence. Our tendency to assign human
goals and values to objects using casual turns of phrase like “the automatic door knows to
open when I walk towards it,” leads to confusion. Aaron is a computer created by Cohen that
draws pictures based on an extensive and complicated rule base. The pictures created are
diverse in style and content, but are they are the work of Aaron or Cohen? Some of the
pictures have titles such as “Mother and Daughter,” but these titles are added after the work
is completed by human observers. Such a title is based in a human observer‟s interpretation
of the content. As such Aaron has no concept of what it has drawn, it is merely following a
rule set. Aaron cannot name or describe its own pictures, and so I would say that the work is
the result of Cohen and not of Aaron. Aaron is the artistic avatar of Cohen, not a separate
entity in its own right. Another computer artist is Ray Kurzwell‟s Cybernetic Poet. This device
can “read” a selection of poets‟ work and reproduce their style. The following piece is the
work of this program.


        MOON CHILD.

        Crazy moon child
        Hide from your coffin
        To spite your doom.

11.    This haiku is, to my mind, an excellent piece of work, yet any inference or meaning I
derive from it comes only from myself, the creator cannot understand its own work.
Anthropomorphism could easily lead us to describe a system as intelligent when it is not, or
to cynically describe an autonomous agent as merely an extension of its creator. (Kurzweil
1999).

12.     Colby, the creator of another program called Parry, has also reported a limited kind of
success in the Turing test with his creation. This variant of the Turing test involved Parry, a
human paranoid, and five psychiatrists. The psychiatrists interviewed, via teletype, either
Parry or the paranoid without being aware of the existence of the simulation. Another group
of psychiatrists were asked to review the transcripts and identify which involved Parry and
which involved real patients. The results of this experiment were striking. None of the original
interviewers realised they were talking to a computer. In ten trials there were five correct
identifications, and five incorrect identifications, no better than chance level. (Boden 1987).


(c) Adam Beautement May 2000                   3
However, significantly, four of the five mistakes involved the patient being mistaken for the
machine. Again the test was not conducted according to the design of Turing.

13.     Despite their limited success these types of program will never achieve intelligence
and, although their output may be impressive, a look at the internal functions of these
programs reveals why. This examination simultaneously reveals the fundamental flaws in
these programs and a fault of the Turing test itself. Both Eliza and Parry work by pattern
matching. When given a sentence they assess the structure and create an answer based on
the input. This leads to the vaguely circular conversations created by these machines. Both
programs include ready made responses sorted into appropriate categorise of replies,
however, Eliza can also generate new responses based on certain input forms. This is a step
forward but quickly leads to its undoing as nonsense sentences are generated. For example
when given the sentence “lets talk about you - not me. Can you think?” Eliza matches it to a
„... you ... me‟ pattern and returns the answer “You like to think I - not you - don‟t you?.” This
is clearly nonsense and immediately reveals the conversationalist to be a machine.

14.     If equipped with enough responses then blind pattern matching could, in theory,
create a conversationalist capable of passing the Turing test, despite clearly not displaying
intelligence. A device based on Parry or Eliza could repeatedly fool a human into believing
that it was not a machine under the conditions of the Turing test. The main stumbling block
in creating such a device is one of knowledge storage. When Copeland said “the design of
large rapidly accessible knowledge stores is a major problem,” he was referring to the fact
that current technology limits the quantity of data available to a program or device. This
limitation means that pattern matchers and the like cannot currently be equipped with
enough (quickly accessible) knowledge to function effectively.

15.     It is possible to look at a theoretical program that does contain such knowledge.
There are a finite number of words in the English language. There is also a finite number of
symbols that can be created using a standard keyboard, even with a user defined graphics
set. It must be possible, given enough time and space, to create every English sentence
using 100 words or less. Following on from this it must be possible to create every possible
conversation in English that uses these sentences. Using one sentence opens up a set of
possible responses. When a response is given, and a reply received, a further response can
be chosen from the stored conversations that start with the sequence of sentences already
used. If all these conversational patterns could be stored and accessed then the resulting
program would surely pass the Turing test. So is it intelligent? Although this super program
would pass the basic Turing test it fails a more intuitive test for intelligence in two main ways.
Firstly it is a specialised program. If it was used for anything other than a conversation via
terminal it would drastically fail. The following quote by Winograd, the creator of SHRDLU
(another Turing test hopeful), exposes these limitations.

        "The A.I. programs of the late sixties and early seventies are much too literal.
        They deal with meaning as if it were a structure to be built up of bricks and
        mortar provided by words, rather than a design to be created based in the
        sketches and hints actually present in the input. This gives them a brittle
        character, able to deal well with tightly specified areas of meaning in an
        artificial formal conversation. They are correspondingly weak in dealing with
        natural utterances, full of bits and fragments, continual (unnoticed) metaphor,
        and references to much less easily formalised areas of knowledge." (Kelly
        1993).



(c) Adam Beautement May 2000                    4
16.     This program is, therefore, not adaptive. Secondly the function of this program, and its
predecessors, Eliza and Parry, reveal it to be nothing but a blind searcher through
predefined conversational forms. The Turing test as it stands is therefore not a test for
intelligence at all but merely a test of conversational and linguistic skills. It would be possible
for a program to pass this test with absolutely no understanding of the words it is discussing,
as the super program example shows. Is it then possible to alter the Turing test in such a
way that it does become a test for intelligence? Examining the major criticisms of the test
and possible changes to counter them it is possible to see how that test may be improved
and also outline some of the criteria for intelligence.

17.     We have so far established that the Turing test does not really test for intelligence but
only linguistic ability. This forms the basis for our first objection, that entities lacking the
power of speech may still display signs of intelligence. Most notably, the higher primates
have demonstrated behaviour indicative of relatively complex goal driven and intentional
mentalities, yet they would fail the Turing test. In the same way, a cognitive machine may fail
the test because it displays responses and behaviour that are non-human, or attempts to
communicate through another medium. This first objection shows that the power of written or
spoken communication is not the optimal criteria for an assessment of intelligence.

18.     Following on from this, the second objection to be voiced deals with the fact that
verbal responses are the only ones required by the test. Yet, as noted above, verbal
responses are not the only means of indicating intelligence. Also, a knowledge of language
is not necessarily grounded in understanding or awareness of the objects the language
refers to. This allows programs to use words who‟s meaning they have no real concept of. If
other responses and sensory abilities were included into the test it would make it possible to
ask the subject of the test to describe objects held before it. Not all conversation includes
direct reference to material objects, however being able to ask questions about real time
input, such as visual stimulus, would allow a new line of inquiry to be pursued that an
intelligent being could surely answer but may cause a machine that lacks intelligence to
falter. This new form of testing immediately discredits the super program as its specialised
nature could not cope with tasks such as describing a view or identifying a red patch when it
can only respond to verbal questioning with pre-set answers. Some items cannot be held up
to visual scrutiny, most notably esoteric concepts like the square root of a number, and an
understanding of those can only be ascertained by suitably probing questions. In this regard
it is not necessary to change Turing‟s test as a verbal response test will never allow all the
responses to situations that indicate intelligence. What this line of objection does reveal is
that the Turing test is not a litmus test. Failure of the test does not immediately rule out
intelligence in the tested entity, and a successful passing of the test does not guarantee the
presence of intelligence.

19.     The third objection to the Turing test as a measure of intelligence concerns the nature
of the implementation of the intelligence in question. As shown above a computer or
program that passes the Turing test is not necessarily displaying intelligence, it could be a
super program such as the theoretical model already described. In this case the model is a
simulation of intelligence, not intelligence itself. If a program is a simulation then it lacks
value as far as determining the nature of intelligence goes. Searle (1980) expresses this
view in the following (rather cynical) way.

        “The idea that computer simulations could be the real thing ought to have
        seemed suspicious in the first place because the computer isn‟t confined to
        simulating mental operations by any means. No one supposes that a computer
        simulation of a rainstorm will leave us all drenched. Why on earth would anyone

(c) Adam Beautement May 2000                     5
        suppose that a computer simulation of understanding actually understood
        anything?”

It is important to make the distinction between a simulated object and an artificial object. A
simulation “...lacks essential features of whatever is being simulated.” (Copeland 1993). An
artificial object has all the same attributes as the thing it duplicates but is created by non-
standard means. The goal of A.I. is to produce intelligence, not simulations of it. The Turing
test allows simulations to pass and thus to be described as intelligent. The problem of
rooting out simulations is solved by the final objection to the Turing test.

20.    The fourth and final objection to be levelled at the Turing test is the so called „Black
Box‟ objection. In the Turing test we are asked to make a judgement about whether or not a
computer thinks based solely on it‟s outward behaviour. Initially this may not seem like a
valid objection as we make judgements about the mentation of fellow humans everyday
based purely on observation. However we do this under the assumption that they are
comprised of the same materials with the same causal properties as ourselves. However if a
friend was revealed as a robot comprised of silicon and wires then our judgement that he or
she was intelligent would immediately be called into doubt. In the same way the outward
appearance of thought can be dispelled by knowledge of the inner workings. Assuming that
either Eliza or Parry did pass the Turing test, one look at the function of the program would
be enough to convince one that the program still does not think. This final objection
invalidates the Turing test as a measure of intelligence as it has now been shown that it is
possible for a program that does not think to pass the test.

21.     Assessing the Turing test has shown that it is not a valid measure of intelligence.
Non-intelligent entities can pass the test and intelligent ones can fail. Intelligence cannot be
assessed by verbal interrogation. Can the test be improved to counter the accusations
thrown at it? The super program used to pass the test was created using knowledge of how
the test works to construct a dedicated solution. Outside of the test the program will quickly
fail. By changing the test, to include any task specified by the interrogator, dedicated
solutions become an impossibility. An open-ended test would be extremely hard to
administer but does instigate what has come to be recognised as a true requirement for
intelligence, that of adaptation. Intelligence is “most generally, the capacity to deal flexibly
and effectively with practical and theoretical problems.” (Blackburn 1994). Any entity hoping
to be described as intelligent must display the ability to deal with any situation it is put in
despite having no prior knowledge. The ability to formulate a rational, if incorrect, response
to unknown or unexpected situations is a prime requirement of intelligence.

22.     Assessing this new criteria for intelligence requires a two-level approach. The
previous Turing test looked only at output criterion. However, in the extended test it is also
necessary to examine the internal or design criterion of the entity under scrutiny. This extra
dimension of assessment will quickly reveal simulations of intelligence by their method of
functioning. Any type of program that provides answers by drawing them from a pre-
programmed data bank will immediately fail the design criterion for the new test. The output
criterion remain much the same as before, to act and converse in a manner indistinguishable
from a human or other intelligent being in the same situation. The situations that the
candidate of the test should be required to respond to should also involve more than verbal
responses. This is because, as pointed out above, verbalisation is a poor indicator of
intelligence, as intelligent beings may be non-verbal.

23.     However the correct design criterion is considerably harder to pin down. The problem
in defining a design criterion for intelligence is that we understand the nature of our own

(c) Adam Beautement May 2000                    6
cognition very poorly. This makes it very hard to distinguish what, in others, matches the
intelligent mental processes we identify in ourselves. It is easy to dismiss false techniques
(such as pattern matching) but how are we to ascertain what is the correct design criterion?
It is excessively anthropocentric to suggest that any intelligent system must be a mirror or a
model of our own cognitive processes for this would immediately exclude any agent,
biological or mechanical, that was not of the human race by direct descent. This heuristic
could well exclude members of the animal kingdom as well.

24.     In looking at the Turing test and its shortcomings we have seen that intelligence
cannot be found by examining purely the superficial output of a system, nor by response to a
single specified test. Also, the lack of ability to verbalise the reasons for a certain behaviour
does not render that behaviour unintelligent per-se. For an agent to be described as
intelligent it must display adaptive behaviour, and have an internal design that does not rely
entirely on an obviously non-intelligent strategy such as blind pattern matching or the
clockwork cycles of reflex and instinct. This differentiation between instinct and the concept
of intelligence is voiced by Guilford in "The nature of human intelligence.":

        Having defined life as "the continuous adjustment of internal relations to external
        relations," Spencer believed that adjustment is achieved by intelligence in man
        and by virtue of instinct in lower animals." (Guilford 1967).

25.      Following this, to achieve a true state of intelligence, A.I. must also adjust to the world
through thought, not simple programmed response. In studying this topic it has become clear
that it is easier to say what is not intelligent than to say what is. By this process of paring
away the dead wood of the unintelligent the true form of intelligence may be revealed.


A NEW TURING TEST?

26.     The goal then of A.I. is to create a massively adaptable system that can be described
as intelligent, not by its responses to verbal questionings or other constrained tests but, by
its adaptive behaviour to novel situations. When attempting to test for intelligence the result
is often more representative of aptitude at intelligence tests, rather than being indicative of a
more fundamental attribute. Behaviour in unconfined situations is a far better indicator of
intelligence and adaptability. Copeland describes an agents inner processes as massively
adaptable if:

        "It can do such things as form plans, analyse situations, deliberate, reason,
        exploit analogies, revise beliefs in the light of experience, weigh up conflicting
        interests, formulate hypotheses and match them against evidence, make
        reasonable decisions based in imperfect information and so forth."

Dennett (1971) says that:

        "an intentional system is one whose behaviour we can explain, predict, and
        control only by ascribing beliefs, goals and rationality to it."

27.   Essentially, an adaptable system is not constrained by its programming, whether
mechanical or biological, and is able to respond to changing situations. Adaptable systems
embody the notion of what agents think, not how they think it. Explanations for the behaviour
and responses of such a system are formulated on a level of intentions and goals, and are
not based on what biochemical and electric processes are taking place. Intrinsic to the

(c) Adam Beautement May 2000                     7
notion of intentional explanations for behaviour is the concept of the agent in question
wanting some goal. An external observer cannot describe this internal activity because the
intended result is not the same as the actual result. Additionally, method is no indicator of
intention (Farr and Moscovici 1984). For example, if the intention is to turn on a light bulb,
while the most common method is to flick a switch, other actions such as tightening or
replacing the bulb, or even banging the wall in an old house may achieve the same result.
This kind of creative behaviour, forming goals - then solutions to them based on current
environmental conditions, defines intelligence more than test results.

28.      Having framed some basic criteria for intelligence we must now turn to the second
part of our question. "What makes an artificial system intelligent." As stated above an
artificial system is merely a duplicate of an existing system created through 'unnatural'
means. This being the case, the same criteria for intelligence apply. However is it possible to
create such a system through artificial means? The notion that the causal nature 1 of our
brain structure is essential to the manifestation of our intelligence has already been voiced.
Current levels of technology lead us to the widely held opinion that digital computers are the
most likely way of implementing an A.I. system, but can the structure of a digital computer
support the same kind of processes as the biochemical form of our brains? The most
famous case against this idea is the Chinese Room argument, voiced by Searle in 1980.

29.     Searle proposed the Chinese room argument in his 1980 paper "Minds, Brains and
Programs." This argument was formulated to try and show that computers, as symbolic
manipulators, could never implement intentionality. The reason behind this is that digital
computers can only operate on symbols, and as such have a mastery of syntax but
absolutely no grasp of semantics. The traditional approach to A.I. as outlined by such
thinkers as Turing and Church would lead us to believe that this lack of semantic ability does
not prevent the creation of intentions. This view is derived from the thesis that for any
algorithm there is a Turing machine that can execute that algorithm. Added to this is the
theory that there is a universal Turing machine that can implement any and all other Turing
machines, and therefore by default all algorithms. If our brains are this universal Turing
machine, then by extracting the algorithms used by our brains through a process of
modelling we can gain an understanding of our cognitive processes. This approach relies on
the view that our brains are merely running a program and also that semantics are mirrored
by the underlying syntax. The Chinese room attempts to show that the one is not intrinsic to
the other. If this is that case then an A.I. cannot be created using current technology. This
essay will now present the Chinese room argument and its major criticisms and attempt to
show that this argument can stand up to its criticisms levelled against it but that it fails to
address the one type of program that, in my opinion, is most likely to yield an A.I.

30.      The basic premise of the Chinese room is this; If a person with no knowledge of
Chinese were to be sealed in a room with a multitude of code books containing Chinese
symbols, and instructions in English on how to manipulate the symbols, he would be able to
receive and sentences in Chinese, process them using the code books, and return sensible
answers, again in Chinese. At no stage does the person in the room ever have to posses a
knowledge of Chinese yet to an external observer the answers produced by the sealed room
would be indistinguishable from those given by a native Chinese speaker. This illustrates
that it is possible to manipulate symbols without ever knowing or understanding what it is
that you are manipulating. The person in the room represents the instantiations of a
program, and performs similar operations on the input and output. Searle argued that this
showed that computers implementing a symbol system can never have understanding of

1
    ie: the nature of the components of the brain and their resultant properties.
(c) Adam Beautement May 2000                                8
what they are manipulating. Obviously this argument generated a lot of opposition from the
traditional A.I. community and several counter arguments were put forward. The most
interesting of these, for the purposes of our discussion, is the so called 'systems reply'.

31.     The systems reply is outlined both by Searle (1980) and Copeland (1993). The thrust
of this counter argument is that while the man in the room (for simplicity referred to as Tom
from here on) does not understand Chinese, the wider system (of which he is a part) does.
Copeland goes on to introduce a new character into the Chinese room here, or rather to
highlight an existing but ignored one. He points out Miss Wong Soo Ling, the voice of the
room and the fruit of Tom's labours. Copeland claims that while Tom does not understand
Chinese, Miss Wong does. Copeland's basis for this claim is the logical flaw in Searle's
argument. In the reasoning behind the Chinese room is the thesis that because no amount
of symbol manipulation will allow Tom to understand Chinese, no amount of symbol
manipulation by Tom will allow the wider system to understand Chinese. This is not a logical
argument. Just because our hearts cannot filter water from blood does not mean that no
other part of our body can. The heart is merely a part of a larger system that is not defined
by the hearts attributes and abilities. However there are two issues to be considered here,
and to mix them up is to lose the value of both the Chinese room and Copeland's reply.

32.     The logical flaw pointed out by Copeland shows that the Chinese room argument
does not prove that symbol manipulators cannot support intentionality. However the systems
reply is in itself flawed. The Chinese room still illustrates that symbol manipulators that use a
system of pre-programmed operations on recognised inputs cannot understand what they
are manipulating. This is demonstrated by the intrinsic difference between syntax and
semantics. Copeland tries to discredit this section of the Chinese room argument in his
critique using his example of Miss Wong. While he is correct to point out that the Chinese
room fails to completely render symbol systems obsolete his argument to support the
systems reply is flawed. It can be easily refuted in two ways, one taken from his own
criticisms of the Turing test.

33.     As discussed earlier in this essay Copeland showed a flaw in the Turing test using the
black box theory. In summary, this theory says that an intelligent system must satisfy both
output criterion and design criterion. Now the Chinese room easily satisfies the output
criterion, its replies are indistinguishable from a native Chinese speaker. However if the
internal design of the Chinese room, and therefore Miss Wong, is examined it reveals a
system no more sophisticated, and no more intelligent than the disgraced Eliza and Parry
programs. So Miss Wong cannot be said to understand Chinese at all.

34.     The second, and more powerful, counter to the systems reply is what I have chosen
to call the translation counter. If we were change the Chinese room slightly from a question
and answer device to a translator then its flaw is quickly revealed. In the new room the code
books have been swapped for Chinese to Russian translation instructions. Now we can ask
the room to translate any Chinese sentence into Russian. Again the outputs would be
indistinguishable from a native Chinese speaker, until we introduce a new type of input. If we
ask both a native speaker and the room to translate a grammatically correct but nonsensical
sentence such as, "spliced into the shy side of darkness, a minute colour of hedonism
crumbled into heaven," then we should receive revealingly different answers. The native
speaker will immediately question the sentence, recognising its nonsensical nature. The
Chinese room will happily translate this, and any other gibberish you ask it to, because its
syntactic expertise is not mirrored (lest by inversion) by its semantic knowledge. This is
because the Chinese room gives responses which are not guided by any internal knowledge.
It has no ability to regulate and monitor it's output through a process of self review. This blind

(c) Adam Beautement May 2000                    9
following of internal rules and programming is exactly the kind of clockwork repetition that
was cited earlier as an example of what is not intelligent behaviour.

35.     Boden (1988) also attacks the Chinese room. She does not refute the claim that the
system has no understanding of Chinese, but instead argues that it may posses
understanding of another kind. Tom, the man in the room, when viewed as an instantiation
of a computer program, does not lack all understanding. Rather he lacks understanding of
the specific subject matter. Tom may not understand Chinese but he may have knowledge of
the nature of conditional IF / THEN relationships. If this is the case then a digital computer,
or symbol system, may still embody understanding. How he understands this or any other
factor is more relevant than whether or not he understands Chinese. This leads to a shift
from "when does a machine begin to understand," to "what things must a machine do in
order to understand." This is because there is no definite cut-off point between
understanding and no understanding, it is not a binary all or nothing state. For all forms of
understanding there must be some underlying mechanism or process that allows
understanding to take place.

36.     This is the mirror argument to Copeland's objection that a lack of understanding of
one subsystem does not rule out lack of understanding for the system. Boden is saying that
lack of system understanding does not prevent subsystems understanding different concepts
relating to their role in the larger system. Boden does not offer any concrete evidence for this
being the case, but merely mentions it as a possibility to consider. It also illustrates the idea
of spiral development, new ideas being formed based on the principals of the old. However
this argument is a rather iterative one. Rather than addressing the issue in hand (that of
system intelligence) Boden obscures it behind another one (the idea that subsystems may
still understand). The point still stands that the Chinese room fails to be intelligent. The
understanding subsystem proposed by Boden is not intelligent in its own right, and could be
replaced by another Chinese room inside the first, thereby starting the whole debate again.
Boden fails to disband the Chinese room, and focuses instead on a new problem that
mirrors the first and adds nothing towards its solution.

37.     So the Chinese room does show that syntax and semantics are separate, and that
knowledge of one does not lead to an awareness of the other. Searle's argument therefore
shows us that systems using a set of pre-programmed symbolic manipulations can never be
said to understand its subject matter. However, as Copeland's critique illustrates, this
statement cannot be generalised to all systems that use symbolic manipulation as part of
their internal design. Boden adds to this, proposing the thought that the subsystem (in this
case the symbol manipulator) may posses understanding of a different kind. This evidence
means that there may still be hope for an A.I. produced using digital computers, as long as it
stays clear of the kind of code book methodology so effectively discredited by the Chinese
Room.

38.    Another basis for dismissing digital computers as the grounds for artificial intelligence
comes from a physical not theoretical point of view. It may be that the unique causal
properties of our brain matter gives rise to our intelligence. If this is that case then only
matter which mimics the physical-chemical properties of our brains can produce thought.
This matter cannot currently be resolved as we do not sufficiently understand the processes
by which our brain supports cognition to positively identify their physical attributes. Also, the
level at which cognitive processes are created is unknown. If we were to reproduce the brain
would we need to mimic the exact biological and material composition of the brain or merely
the electrochemical signals and the pattern of neurone interactions? There may be other
higher or lower levels where cognition truly resides but until we understand out own cognition

(c) Adam Beautement May 2000                    10
we cannot say whether or not these attributes, physical and procedural, are present or
absent from the materials and designs in digital computers and other artificial systems.

39.    How then to escape from the Chinese room? To free ourselves from its confines we
must satisfy ourselves that a machine created using computational processes can display
adaptive intelligence. Turing quotes Professor Jefferson's Lister Oration for 1949 on this
topic of machine consciousness.

        "Not until a machine can write a sonnet or compose a concerto because of
        thoughts and emotions felt, and not by the chance fall of symbols, could we
        agree that machine equals brain - that is, not only write it but know that it had
        written it. No mechanism could feel (and not merely artificial signal, an easy
        contrivance) pleasure at its successes, grief when its valves fuse, be warmed
        by flattery, be made miserable by its mistakes, be charmed by sex, be angry or
        depressed when it cannot get what it want."

40.     This seems to paint a fairly pessimistic picture of the situation, for how can we
possibly make machines that are kind, resourceful, beautiful, friendly, have initiative, a sense
of humour and other such attributes when we understand so little about what creates these
attributes in ourselves. Yet although we cannot create a formula for these events and
emotions we recognise them in ourselves and in others. This is the key to the riddle of
intelligence, consciousness and all the other problems that plague A.I. Professor Jefferson
briefly touches on it in his speech. To extract the relevant phrase:

        " - that is: not only [to] write it but know that it had written it."

Searle (1980) includes, in a list of things widely believed to be impossible for machines, the
idea that machines "be the subject of its own thought." Yet this is exactly what is needed if
ever A.I. is to be created. Current attempts at A.I. revolve around creating a program that
can understand its subject matter, whether that is the English language, or a game of chess.
Yet all that has resulted are programs that tend to rely on processing power and large
databases. Moreover, when they are moved outside of their field of expertise they can no
longer function. It is not possible to program an intelligent system or agent because by the
very act of programming intelligence you are creating a finite system that cannot step outside
the boundaries of it's program. At best a simulation of intelligence will be created, nothing
more.

41.      Turing proposed what I believe to be the most appropriate method of creating an
artificial intelligence. He suggests that the best way to equip a computer to pass his test
would be to provide it "with the best sense organs money can buy and subject it to an
appropriate course of education." This means creating a program that can be taught and
teaching it, or letting it learn, all it needs to know. It is significant to consider at this point that
a human baby would fail the Turing test (flawed as it is) yet no one would expect it pass, or
dismiss it as unintelligent because it fails. Instead of creating a fully fledged English speaker,
why not create a program that can learn English, or any other language for that matter? An
agent programmed to understand English will only ever be able to utilise that part of the
specific language that it has been instructed to use. However, an agent that can learn
language can both add to its existing knowledge of language, and learn new languages if
necessary. Searle says, when arguing that digital computers cannot support intelligence,
that:



(c) Adam Beautement May 2000                         11
        "mental states and events are literally a product of the operation of the brain, but
        the program is not in that way a product of the computer."

This is borne out by his Chinese room argument, a program created outside a computer and
given to it only as a system of symbols and operators will never have any meaning to the
computer, and will never give rise to understanding. Any intelligence shown is only a
reflection of the programmer. The program of this type will always be limited by what it was
given to do. However if the program is a product of the computer, then perhaps a new
avenue of attack to the goal of a true A.I. has been opened up. Manipulations of predefined
symbols have been shown to not provide understanding. However if the symbols being
manipulated by a system have been defined by the system then understanding becomes
inherent. The system using the meta-symbols has then created them itself based on
whatever input it has received. This will also create associations between symbols. Similar
symbols, or ones that appear together, will be inextricably linked. Additionally if the system
can create its own symbols then a new level of consciousness can be obtained.

42.      Consider the example of the tiger. The symbol of the tiger may both refer to an object
in the vision of the system, but also to a group of other symbols such as the icons for stripy,
orange, black, cat etc. While it is possible to program this into a system, the initial conditions
can never be changed, and consciousness will not arise. However if a system can
manipulate its own symbols then a tiger may become associated with fear and danger or
beauty and power, depending on the systems experience and knowledge of the tiger. Here
we see the beginnings of personality. This is the difference between a native Chinese
speaker, and the Chinese room. The speaker learnt first-hand the meanings and appropriate
utilisation of the symbols he or she uses to answer questions and hold conversations. The
Chinese room did not and so cannot monitor its output as it has only predefined
manipulations to work with. However, the native speaker can monitor such things and will,
therefore, query nonsensical sentences as s(he) will have an idea of what it is wrong.

43.    Therefore, an intentional system will be able to add to it's vocabulary and concepts
when it comes across new words. For example the new symbol 'ubiquitous' could be paired
with the existing symbol for 'everywhere'. Additionally any symbols added will have meaning
only to those that defined them. This is part of the reason why you cannot program
understanding. There are no universal symbols understood by all, not at a cognitive level.
Instead any symbolic representations that may be used will be personalised and
incomprehensible to others. Whether these self manipulations take place consciously or
unconsciously is another factor to consider. Obviously the human equivalent of such
operations take place subconsciously, but whether there will be a distinction in the
consciousness of a machine is an unknown factor.

44.     There is however an immediate problem with this new and hopeful line of thought,
which can be summarised in the following way. To teach and alter itself (for example to
derive new symbols) a program must have itself as the focus of its code and thought. Turing
described this notion in the following way, "by observing the results of its own behaviour it
can modify its own programmes as to achieve some purpose more effectively." This is all
very well but for a program to have itself as its subject requires the program to have a focus
and not just be a slave to the input decided by the operator of the program. This requires the
program to have intention. Intentional states direct the mind at an object, they become an
intrinsic focus on something, and relate the mind to the world (Boden 1988). So to create a
program that can have itself as its focus, we must first create a program with intention. This
is a point also made by Turing, who said, "the claim that a machine cannot be the subject of
its own thought can of course only be answered if it can be shown that the machine has

(c) Adam Beautement May 2000                     12
some thought with some subject matter." This means that before the concept of self-aware
machine can be debated, it is necessary to first have an aware machine. Thought and
consciousness do not necessarily go hand in hand (consciousness in this instance implying
a degree of self-awareness).

45.     One possible method of beginning the process of awareness is to provide a system
with unrestrained access to the real world. An entity that only has input given to it by a
controller will be limited in the same way as a system with programmed abilities. Both of
these agents will be extensions of their creators, not separate beings. If a program or
computer was given an appropriate set of sense organs then it could sample anything it
wants from the world. A vast storage system would be required for such an entity, as it would
be receiving huge amounts of data every second. To cope with this the system would also
need to be able to focus only one section of its input or be overwhelmed. However if as soon
as you create a heuristic for attention, i.e. the loudest source of sound, you are confining the
system. To create such a device would require it to be equipped with a similar set of basic
instincts and responses as a human new-born baby. Whether this type of system is possible
or not is a subject outside the scope of this essay. If it is possible to create such a device
there is no current evidence to suggest that a digital computer could not be used for its
implementation.


CONCLUSIONS

46.     In conclusion, this essay has demonstrated that intelligence is dependant on more
than outward appearance, and that there is no logical reason why digital computers could
not support an artificial intelligence. In some respects this has been a work of negation. Its
concrete claims are addressed towards what is not true rather than what is. It is not true that
the Turing test is a true measure of intelligence, as shown by the criticisms levelled at it.
Some ideas as to what may be required of an intelligent being have been proposed, namely
intentionality, an ability to create its own symbols, adaptability (and to be self-adapting) and
understanding, but no replacement or modification of the Turing test can be suggested that
will be a concrete test of intelligence.

47.     The Chinese room does not demonstrate that symbol manipulators cannot be part of
a system that embodies understanding. Intentionality is an attribute of the system, not its
components. The brain does not have intentionality, yet it is part of a system that most
definitely has. This is why the refutal of Searle's claim that symbol systems cannot support
understanding is so important. By effectively dismissing this idea it keeps alive the possibility
that symbol manipulators can be involved in a larger system that does have understanding
and intentionality.

48.     While this essay has shown that there is no reason to think that digital systems
cannot support understanding, it has not been able to confidently state that they can.
Instead, a possible method of creating an intelligent system is presented, putting forward the
notion that a program that can have itself as its focus will be most likely to produce
intelligence. The greatest flaw in Searle's proposal is that he says that the system is the
bearer of intelligence and understanding and not the causal basis of it.

49.    User defined systems will never hold true understanding as they have no ability to
create their own symbols and to form associations between objects, and between input from
the world and internal concepts. If a user tries to write intelligence into a system it is trying to
make the system a bearer of intelligence. A system that can affect itself may gain

(c) Adam Beautement May 2000                      13
understanding as it grows towards intelligence. Instead of trying to directly define
intelligence, the user has implemented a system that holds the possibility of intelligence.
Intelligence is not hardwired into the system but is an emergent property that arises
from the interaction of its various (possibly self-defined) subsystems.

50.     Current A.I. approaches have failed because they tried to create a single intelligent
system. Intelligent systems will not be created, they will evolve. The key to this approach is
to create a set of self-adaptive subsystems that interact to manifest intelligence at a higher
level than any of the individual systems are operating at. This idea of interaction also helps
us asses the intelligence divide more objectively. There is still a grey area, maggots and
slugs are clearly on one side due to their spartan interactions, both with each other and their
environment. Chimpanzees are on the intelligent side, due to their complex social behaviour.
Our own intelligence arises from the interaction between the different areas and layers of the
brain, between the different hemispheres, and between ourselves and other humans.

51.     There is a precedent for this nature of thought to be found in both psychology and
common phrases. Individuals reared in isolation are always less able than those reared in a
social group. Without the necessary level of interaction they fail to develop to the same level
of intelligence as others. We need social interaction to fulfil our intellectual potential.



Adam Beautement
Nottingham
May 2000




(c) Adam Beautement May 2000                   14